adding optimizations

[rachadele]

adding optimizations

[iprada]

Hi Rachel!

It is great to see people using bulkDGD. We are happy to help!

It would be useful if you describe a bit your experimental set-up (How many samples? which tissue?) and what you are trying to achieve. We can take it from there.

If you do not want to describe your set-up publicly, you can find my email here: https://di.ku.dk/english/staff/?pure=en/persons/525785

best,

Iñigo

[rachadele]

Hi Iñigo,

thanks for responding :- ) I am using breast epithelium RNA-seq data from this study: GSE141828. I am only focusing on the "susceptible" breast tissue samples (for now).

After convering HGNC to ENSEMBL, dropping non-uniquely mappted genes, and, preprocessing samples using ioutil.preprocess_samples, I have a data frame of 7 samples × 16883 genes.

[iprada]

I see. We are working on providing loss curves as a dataframe, but that might take some time.

In the meantime, I would plot the loss curves (x axis = epoch; y axis= loss) to see how they behave. From what you send loss does not look crazy high.

best,

Iñigo

[rachadele]

Hi Iñigo,

Thanks for your suggestions. I've included the plots for the loss curves here.
As you can see the loss is decreasing, I'm just a little confused because in other models I've used the loss is always <1. So I'm not sure if a loss of 14 is acceptable for DGD.

[rachadele]

another important question: should I be providing DGD with raw or normalized counts?

[iprada]

Hi!

DGD should be provided with raw counts. It takes care of the normalization internally (mean scaling). Loss curves look fine and the high numbers you observe is probably the GMM penalty, but I will increase a bit the number of epoch to ensure they converge. In think if you sept op1 epochs to 20 and opt2 to 100 the will look stable.

best,

Iñigo

[rachadele]

Thanks for your help! :- )

I'm also wondering if you can help me generate a figure similar to figure 3B of your paper: .

I assume I have to somehow save the outputs of the probability mass function from the decoder class, but I'm unsure if there's an easy way to do this.

[iprada]

Hi,

Definetely. The decoder has a Negative-binomial layer with a log_prob method implemented. You can use that to calculate the probability of your samples. You simply pass your data through the log_prob method and I should get sample probabilities.

good luck!

Iñigo

[rachadele]

Thank you so much! :)
what are the "scaling factors" for this function? how can I compute them? I tried passing the r_values as in the perform_dea function, but it doesn't accept them.

I do have another question– is there a way to figure out which components correspond to which tissue?
All of our (breast) samples were assigned to component 28, except for one, which was assigned to component 23.

[iprada]

sorry this took so long. Component 28 is the breast component, good news. Component 23 is a brain component. We should publish this data but we need to think how to. In the meantime, if you email me, I can share the data. You can find my email here

https://di.ku.dk/english/staff/?pure=en/persons/525785